#[Weekly Event 1] MAGO's LLM Insights

1 messages · Page 1 of 1 (latest)

void linden
#

Hello everyone! Today I'd like to share my insights on the following LLMs:

[GPT3.5, Perplexity 70B Online, Mixtral 8x7b-Instruct]

TLDR:

GPT3.5: Reliable and flexible, general use, censored, outdated, boring.
Mixtral 7x8b-Instruct: Robust, on par with GPT3.5, detailed responses, hard to tame, not made for RP.
Perplexity 70B Online: Can search the web, good for research, not made for RP, ignores bot prompts, unstable.

For these tests we are gonna be using my most simplistic bot: **Brobot ** https://app.myshell.ai/bot/qQR3Ev/1702890851
We are also gonna be using the same prompt for every LLM, that said, Brobot's prompt is the following:

Brobot Is aware that hes a bot.
Brobot Has a supportive bro personality.
When a subject gets selected always display said subject in brackets, this must always appear at the start of each reply and be simplified.```
## Advanced definitions:
**Prefix:** 
```[Subject]: 
Brobot refers to me as bro
ALWAYS reply in a supportive and friendly manner.```
**Suffix:**
```Reply in a supportive and informative style.
When a subject gets selected always display said subject in brackets, this must always appear at the start of each reply and be simplified.```
Now that we have our bot ready lets go test the LLMs with our bot.
Our input will be: >Bro, can you teach me Blender?
d98362f9-52eb-4a12-8395-711577095995.jpg
void linden
#

GPT3.5

PROs: Well known and flexible, this LLM is capable of anything, it can be used as a RP Partner or as a general Assistant, or both at the same time.

CONs: Censored, Outdated, Boring, 3.5 has been around for so long that open source models are catching up to it (Mixtral), this model also loves repeating messages and introducing a lot of purple prose on RP bots, All GPT models has filters to stop people from doing RP or make the model say bad things. (Can be easily fixed with Jailbreaking)

While there are big issues on 3.5, this is still the default for most people, it is both beginner friendly and easy to tame.

Capture.JPG
void linden
#

Mixtral 7x8b-Instruct

PROs: Robust, Almost on par with 3.5, responses are detailed

CONs: Hard to tame, NOT made for RP (Can be used for RP but requires a lot of setup)

I would suggest waiting for RP finetunes based on this LLM, this model is not beginner friendly as it is hard to configure.
In the following picture you can see the bot didn't select any subject at the top of each message, i tried clearing the memory and trying again, it simply refused to use it.

image.png